Abstract

Deep learning-based computational pathology approaches are becoming increasingly prominent in histopathology image analysis. However, these images typically come with drawbacks that hamper automatic analysis, which include: labeled sample scarcity or the extremely large size of the images (ranging from  to  pixels). Nonetheless, they have proven to be a powerful tool for diagnosis and risk prevention. One such prevention is reducing the number of patients who undergo surgeries that do not benefit them. This study develops a pipeline for predicting sentinel lymph node (SLN) metastasis non-invasively from digitised Whole Slide Images (WSI) of primary melanoma tumours. Furthermore, we combine the use of a weakly supervised architecture with self-supervised contrastive pre-training. We experimentally demonstrate that 1) the use of self-attention improves sentinel lymph node status prediction and 2) self-supervised contrastive learning improves the quality of the learned representations compared to a standard ImageNet pre-training, which boosts the model's performance.